Implementing Genotype-Dependent Parameters in PBPK

From Genotype Effect to a PBPK Parameter

A genotype label does not change a simulated concentration by itself. It must be translated into a model input, such as enzyme abundance or pathway-specific intrinsic clearance, at the appropriate place in the PBPK model. The choice of translation matters: a fixed value for each genotype, a proportional activity-score adjustment, and a distribution conditional on genotype encode different assumptions about biological variation.

The genotype and its functional interpretation are inputs to this implementation step. Phenoconversion or another clinical modifier may mean that genotype-based activity is not the final effective activity; apply such modifiers consistently with the model’s stated assumptions rather than silently changing the genotype-to-parameter mapping. The introductory material on genetic variation as a PBPK input provides the context for treating genotype as a model covariate.

A useful mechanistic starting point is the enzyme rate expression

Here \(v_g\) is the metabolic rate, \(C\) is the relevant substrate concentration, \(E_g\) is enzyme abundance, \(k_{\mathrm{cat},g}\) is catalytic turnover per amount of enzyme, and \(K_{m,g}\) is the Michaelis constant. At concentrations well below \(K_{m,g}\), the rate is approximately \(E_g k_{\mathrm{cat},g}C/K_{m,g}\), so enzyme-specific intrinsic clearance is proportional to \(E_g k_{\mathrm{cat},g}/K_{m,g}\). This explains why a genotype effect on abundance can affect clearance, but it does not imply that every genotype effect should be implemented by scaling abundance. A variant can instead alter catalytic function or substrate affinity.

Before choosing a method, specify the biological quantity affected, its model location, and its units. For example, if the PBPK model already calculates pathway clearance from enzyme abundance, modify abundance rather than also multiplying the resulting clearance by the same genotype factor. If the model represents the pathway directly by \(CL_{\mathrm{int}}\), apply the effect there. The implementation should reflect one coherent representation of the effect.

Three Parameterization Strategies

Categorical parameter assignments

A categorical assignment uses a lookup table: each genotype-derived group maps to a fixed parameter value or multiplier. This is appropriate when the evidence supports group-specific values and the model is intended to represent a typical subject in each group. It is transparent and straightforward to audit, but it does not represent variation among subjects within a group unless additional variability is included elsewhere.

In this expression, \(\theta_g\) is the parameter for group \(g\), \(\theta_{\mathrm{ref}}\) is the reference-group value, and \(r_g\) is a group-specific multiplier. A lookup table may specify \(\theta_g\) directly instead. The reference and all assigned values must use the same parameter definition and units.

Activity-score scaling

An activity-score implementation maps a genotype-derived score to a continuous multiplier. If a reference score \(A_{\mathrm{ref}}\) corresponds to a calibrated reference parameter, proportional scaling gives

This rule assumes that the score is on a meaningful ratio scale and that the parameter is proportional to the score over the range being used. It is not justified merely because a score is numeric. For example, proportional scaling may be a reasonable approximation for enzyme abundance or low-concentration intrinsic clearance when allele activity contributions combine approximately linearly. It may be unsuitable if the score is ordinal, if the parameter has a nonzero residual floor, or if saturation, regulation, or other pathways make the relationship nonlinear. Scores and reference values must also be defined consistently; a ratio to an arbitrary reference can otherwise produce a misleading multiplier.

Genotype-conditional distributions

A distributional implementation assigns a parameter distribution conditional on genotype. It represents residual differences among subjects with the same genotype, such as variability in enzyme abundance not explained by genotype. A common positive-valued model is lognormal:

Here \(m_g\) is the conditional median and \(\mathrm{GSD}_g\) is the geometric standard deviation. This parameterization is useful for positive quantities and multiplicative variability. It is important to distinguish the median \(m_g\) from the arithmetic mean, which is \(m_g\exp(\sigma_g^2/2)\). Use the distribution that matches the evidence and the model’s intended variability; do not add a distribution simply to make an input appear more realistic.

For a population mixture, the unconditional density can be written as

where \(\pi_g\) is the proportion in group \(g\). This is a mixture of genotype-conditional distributions, not a single distribution shared by all subjects. For one virtual subject whose genotype is already specified, use that subject’s conditional component; the mixture weights are not used to reassign the subject’s genotype.

Worked Examples

Why Activity-Score Scaling Can Be Reasonable

The proportional rule follows directly when the relevant biological quantity is proportional to activity score. Suppose an enzyme’s effective activity is \(A_g\), while all other determinants of its low-concentration intrinsic clearance are unchanged across groups. If \(CL_{\mathrm{int},g}=kA_g\) for a common constant \(k\), then for a reference group \(CL_{\mathrm{int,ref}}=kA_{\mathrm{ref}}\). Dividing the first relation by the second gives

The cancellation of \(k\) is the basis for scaling. It depends on a shared proportionality constant, comparable assay or calibration conditions, and no important genotype-dependent change in the other determinants represented by \(k\). If genotype changes \(K_m\), catalytic turnover, or the relationship between abundance and functional activity, the proportionality may fail. A direct genotype-specific parameter estimate or a mechanistic enzyme model may then be preferable.

Implementation Checks and Choice of Approach

A PBPK implementation should be checked at the parameter level before interpreting exposure predictions. Confirm that the subject’s genotype-derived group maps to exactly one intended rule, that the assigned parameter is in the expected units, and that the effect is placed at the correct tissue and pathway. Then compare the resulting parameter against reference and group-specific values, including boundary cases such as the lowest activity score or a group assigned zero activity. A value of zero is a strong biological assumption, not a harmless computational default.

These approaches can be combined only when their roles are explicit. For example, a genotype-specific median with a conditional lognormal distribution can represent group differences plus within-group variability. Applying a genotype multiplier to an already genotype-specific median would count the group effect twice unless the multiplier has a separate, justified purpose. Likewise, scaling enzyme abundance and then scaling the clearance calculated from that same abundance duplicates the effect in a model that derives clearance from abundance.

Diagnostics include tracing a virtual subject’s genotype through the mapping to the final parameter, testing that reference genotypes recover the intended reference value, and checking that changing only genotype changes only the intended inputs. Implausible discontinuities between adjacent activity scores may signal a mismatched lookup or scale. Implausibly identical parameters across genotype groups may indicate that the mapping was not applied. When a distribution is used, inspect its median, spread, and tails in the parameter’s units; a positive-valued distribution can still generate biologically implausible values.

No one method is universally best. Categorical assignments are easiest to explain and audit when evidence supports discrete group values. Score scaling is compact and can interpolate across scores, but can create false precision if the activity scale is poorly calibrated. Conditional distributions express residual heterogeneity, but require distributional assumptions and evidence for their spread. In all cases, parameter-level plausibility is necessary but not sufficient: the model’s predicted concentrations also depend on other pathways, physiology, and the PBPK structure.

Common Mistake

Another frequent error is to treat activity scores as though they were automatically proportional to every enzyme parameter. Before scaling, state what the score measures, what reference it uses, and why the target parameter should change in proportion. If those conditions are not defensible, use a supported categorical mapping or a mechanistic parameterization instead.